According to The Decoder, researchers have developed a new benchmark called Taste-Bench. It is meant to measure how well AI agents make strategic decisions on long-term tasks. The term refers to a kind of "taste", a sense of which path is the better one in the long run.

What the test shows

The result is sobering. The best model tested, GPT-5.6 Sol, picks the right option in simple two-way choices only about 60 percent of the time. With a choice between just two possibilities, that is not a confident score. Pure guessing would get about half of the cases right.

Another finding stands out: thinking for longer changes little. Many current models can put more computing time into reasoning before they answer. For math problems or programming tasks, this often helps. According to the study, however, this lever seems to do little for strategic judgment.

More on this topic:

Why this matters in practice

AI agents are systems that do not just answer individual questions but work independently over many steps. They plan, choose tools and set intermediate goals. On such long-term tasks, small wrong decisions add up. Anyone who turns in an unfavorable direction early on loses a lot of time later or has to start over completely.

That is why a yardstick like Taste-Bench is useful. Many tests check whether a task is solved correctly. Here the question is whether an agent recognizes what is worth pursuing in the first place. That is harder to measure, because there is rarely a clear, immediately verifiable solution.

For users, this means: important strategic decisions should still be reviewed or made by people. Agents can take over routine work and partial steps. Whether they should be allowed to set a direction is something to weigh carefully in critical projects. This is a cautious conclusion drawn from a single benchmark. It shows a weakness, but not a complete picture of everything today's models can do.